Performance improvements

[alexander-yakushev]

Hi, here are a few improvements I discovered while working on Metabase.

The first commit addresses the following issue. Binding dynamic variables is comparatively quite expensive. honey.sql/format binds a lot of dynvars, but most of them usually don't receive meaningful values. A different approach is possible – prepare a bindings map that would contain only the vars with values that actually have changed, and pass that to with-bindings macro. Here are ad-hoc benchmark results obtained from Metabase tests:

Before
Time per call: 228.36 ms   Alloc per call: 559,342,831b
Time per call: 224.91 ms   Alloc per call: 559,342,813b
Time per call: 228.88 ms   Alloc per call: 559,342,813b

After
Time per call: 205.92 ms   Alloc per call: 452,848,231b
Time per call: 202.64 ms   Alloc per call: 452,848,232b
Time per call: 205.16 ms   Alloc per call: 452,848,229b

So, -20% allocations and -10% execution time.

The second commit ensures that xs passed to format-values is a vector because this function traverses xs multiple times using reduce, and the latter is inefficient for arbitrary seqs. The commit also contains a few minor changes that swap lazy seq processing for transducers.

[seancorfield]

This seems to break bb and cljs compatibility.